AI with guardrails – built into your product
Not a gimmick but features that take work off your plate: researched content drafts, automatic image descriptions, data entry at the push of a button or a smart search. Proven in production – with verifiable sources and human approval.
- 3 systems in production
- Claude API · OpenRouter
- Structured output
- Human approval
What you get
Use-case analysis
We find the places where AI genuinely saves time – and are honest about the ones where it doesn’t.
Content pipelines
Research with real sources and drafts straight into the CMS. Your editors review and publish.
Vision & media
Automatic alt text, image descriptions and classification – right at upload.
Data intake
Unstructured sources in, checked structured records out – hours of manual entry become one review click.
Search & assistants
A smart search or an assistant that answers based on your own content.
Guardrails & control
Schema validation, required sources, per-generation cost tracking and approval workflows are built in from the start.
Guardrails, not a black box
AI models are powerful but not infallible. That’s why every generation in my systems runs through fixed guardrails: the model answers into a predefined schema instead of free text, facts need a verifiable source, every call is logged with its cost – and only what a human has approved goes live.
- Schema-enforced, structured output
- Web research with citable sources
- Cost tracking per generation
- Nothing goes live without human approval
Where AI earns its keep
Three AI subsystems run in production on the Experten platform: a content pipeline researches keywords and sources and puts finished article drafts into the CMS. An intake assistant turns a Google Maps URL into a checked profile draft. And a vision model writes accessible alt text at upload. The same building blocks transfer to your product.
- SEO content pipeline with citations
- Profile intake from a single URL
- Vision alt text at media upload
- Transferable to shops, directories and knowledge bases
Does this fit your project?
Write to me about what you need. You’ll get a reply within 24 hours and, after a short call, a fixed quote.
How it works
Find the use case
Which task costs the most time today? That’s where we start – with a clear goal for how success is measured.
Prototype on real data
A quick prototype on your real content shows how good the results actually are.
Integration & guardrails
Built into your CMS or product, with schema, source checks and cost tracking.
Approval & operation
An approval workflow for your team, then monitoring quality and cost in production.
Technologies
- Claude API
- OpenRouter
- JSON Schema
- Vision
- Web grounding
- Payload CMS
- Next.js
- PostgreSQL
In practice
Frequently asked questions
Will my data be used for training?
I work through API access from providers that don’t use API data to train their models, and send only the data a task actually needs. Which providers and data are involved is agreed together before we start – with the GDPR in mind.
What does it cost to run?
Model costs depend on usage and model and are logged per generation – so you always see what a feature costs. Up front I estimate the running costs based on your expected usage.
Will AI replace my editors?
No, it takes load off them. The AI delivers drafts and suggestions; the decision stays with people. That’s exactly how my systems are built: nothing goes live without approval.
Can AI be added to my existing website?
Often yes – as a feature in the CMS, as search or as background automation. Whether and how depends on your system; I’ll look into it after your request.
Your question isn’t here? Write to me – reply within 24 hours
Goes well with
Let's plant your project
Tell me about your idea – I reply within 24 hours.